A fan fault diagnosis method fusing graph domain coding and residual neural network

Abstract

The application discloses a fan fault diagnosis method fusing graph domain coding and a residual neural network, and belongs to the technical field of fan fault diagnosis. The method comprises the following steps: collecting vibration signals of a fan under multiple states; dividing each vibration signal into multiple one-dimensional time domain sub-signals; coding each one-dimensional time domain sub-signal into a two-dimensional graph domain by using a fine recursive graph to reflect unique internal structural features of the signal; proportionally dividing graph domain features into a training set and a test set; learning a nonlinear relationship between graph domain textures and faults from the training set by using a residual neural network, and then constructing an intelligent fault diagnosis model; and then inputting the test set into the trained residual neural network to complete fault diagnosis of the fan. The application can accurately identify the fan fault type, provides a strong basis for operation and maintenance of the fan, and ensures reliable and stable operation of the wind turbine generator.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine fault diagnosis technology, and more specifically, relates to a wind turbine fault diagnosis method that integrates graph domain coding and residual neural networks. Background Technology

[0002] Wind turbines are an indispensable part of the wind power generation industry. However, due to the harsh working environment, their mechanical transmission, electrical, and other systems inevitably experience failures, leading to accidents and economic losses. Therefore, accurate fault diagnosis of wind turbines can be used to develop reasonable maintenance plans, extend the service life of wind turbines, ensure personal safety, and create greater power generation benefits.

[0003] Numerous studies related to wind turbine fault diagnosis have been published, with vibration signal analysis-based diagnostic techniques receiving increasing attention. By analyzing vibration signals to extract fault-sensitive features and combining this with pattern recognition methods, intelligent fault diagnosis of wind turbines can be effectively achieved. However, the non-stationarity and complexity of vibration signals make feature extraction more challenging. To address this, many advanced signal processing techniques have been proposed and applied to analyze non-stationary and complex vibration signals for fault diagnosis, such as wavelet packet transform, empirical mode decomposition, ensemble empirical mode decomposition, and variational mode decomposition. However, environmental and equipment noise inevitably mixes into the acquired signals, suppressing fault frequencies and significantly limiting the application of these methods. Furthermore, these methods cannot intuitively and effectively capture the dynamic changes of vibration signals. Summary of the Invention

[0004] To address the shortcomings of existing diagnostic technologies, such as the inability to intuitively and effectively characterize the dynamic evolution of faults, a wind turbine fault diagnosis method integrating graph domain coding and residual neural networks is proposed. This method achieves high-efficiency and high-precision diagnosis, providing strong guidance for ensuring the reliable, safe, and economical operation of wind farms.

[0005] To achieve the above objectives, this invention provides a wind turbine fault diagnosis method that integrates graph domain coding and residual neural networks, comprising:

[0006] (1) Collect vibration signals of the fan under various conditions;

[0007] (2) Divide each vibration signal into multiple sub-signals;

[0008] (3) Use a fine recursive graph to encode the one-dimensional time-domain sub-signal corresponding to each vibration signal into a two-dimensional texture map domain, and obtain the two-dimensional texture map domain corresponding to each vibration signal to reflect the structural features of the fault hidden in the vibration signal.

[0009] (4) Divide the two-dimensional texture map features into training set and validation set according to the proportion;

[0010] (5) Use residual neural networks to learn the nonlinear relationship between two-dimensional texture map domain and faults from the training set, and then construct an intelligent fault diagnosis model;

[0011] (6) Input the test set into the trained residual neural network to complete the intelligent fault diagnosis of the wind turbine.

[0012] In some optional implementations, the vibration signal is historical data of different operating states recorded by the wind turbine and wind farm in operation, and labels are set according to the state category, with the number of labels equal to the number of state categories.

[0013] In some alternative implementations, step (2) includes:

[0014] Each vibration signal is divided into non-overlapping sub-signals of equal length.

[0015] In some alternative implementations, step (3) includes:

[0016] Phase space reconstruction is performed on each sub-signal corresponding to each vibration signal to obtain the embedded sample sequence in the phase space corresponding to each vibration signal;

[0017] The similarity distance between all pairs of embedded sample sequences in the phase space is calculated using Euclidean distance.

[0018] The trend information of phase space vectors is calculated. Based on the similarity distance and trend information, a fine recursive matrix is ​​calculated and encoded into a two-dimensional texture map domain.

[0019] In some alternative implementations, step (4) includes:

[0020] The two-dimensional texture domain is divided into two parts, with 80% of the data randomly selected as the training set and the remaining 20% ​​as the test set.

[0021] In some alternative implementations, step (5) includes:

[0022] Transform the size of the 2D texture map domain and use it as input to the ResNet18 residual neural network;

[0023] Input the training set, select cross-entropy as the loss function, and train the ResNet18 network using gradient descent.

[0024] In some optional implementations, ResNet18 includes 17 convolutional layers, 2 pooling layers, and 1 fully connected layer. Except for the first convolutional layer, the remaining 16 convolutional layers are divided into 4 groups of residual convolutional blocks. Each group of convolutional blocks includes 2 residual units, and each residual unit includes 2 convolutional layers connected in one step. An identity connection is established between the input and output of the 2 convolutional layers. The ReLU activation function is used to perform a nonlinear transformation on each convolutional output y. The number of outputs of the last fully connected layer is modified to the number of wind turbine fault types. The above layers are connected in sequence. The structure of the ResNet18 neural network can be represented as: conv1, max pooling layer, 4 groups of residual convolutional blocks, average pooling layer, and fully connected layer.

[0025] In some optional implementations, the test set is input into the trained ResNet18 neural network, and the network output is compared with the actual label category of the wind turbine to obtain the diagnostic accuracy, thus completing the intelligent fault diagnosis of the wind turbine.

[0026] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0027] This invention can be used for wind turbine fault diagnosis. To reveal the dynamic characteristics of internal fault changes in signals, a refined recursive graph is innovatively proposed. This graph visualizes the changes in wind turbine state in the form of a two-dimensional topological structure and is organically integrated with a residual neural network to further learn the relationship between faults and the two-dimensional graph domain, thus proposing a wind turbine fault diagnosis method. In this method, the refined recursive graph has the advantage of clearly reflecting the physical behavior related to wind turbine faults. ResNet18, as a deep neural network, has strong nonlinear fitting capabilities and can deeply explore the structural differences between different states in the two-dimensional graph domain. Therefore, integrating the refined recursive graph and ResNet18 to provide a quantitative wind turbine fault diagnosis method can improve the accuracy and interpretability of equipment faults. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the implementation of a wind turbine fault diagnosis method that integrates graph domain coding and residual neural networks, as provided in an embodiment of the present invention.

[0029] Figure 2 This is a flowchart of a refined recursive graph implementation provided by an embodiment of the present invention;

[0030] Figure 3 This is a fine recursive graph of normal and fault conditions provided in an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0032] The purpose of this invention is to provide a new assessment method for wind turbine fault diagnosis, to provide a method for visualizing fault evolution, to improve assessment accuracy, and to solve the problems of traditional methods that are difficult to intuitively and effectively characterize the dynamic evolution of faults.

[0033] This invention is achieved using the following technical solution: Figure 1 As shown in the figure, an embodiment of the present invention provides a wind turbine fault diagnosis method that integrates graph domain coding and residual neural networks. The overall diagnosis process can be divided into two stages: the first stage is to use a fine recursive graph to realize two-dimensional graph domain coding of a one-dimensional signal, revealing the topological structure features of the fault within the signal; the second stage is to use ResNet18 to learn the corresponding features and perform fault diagnosis, including the following steps:

[0034] Step 1: First, collect historical vibration signals and operating data of the wind turbines in the wind farm;

[0035] Taking wind turbine gearbox as an example, the collected health status data of wind turbine gearbox includes nine typical states such as normal state, broken tooth, base crack, tooth damage, gearbox failure, and generator failure.

[0036] Step 2: Divide the acquired vibration signal into non-overlapping sub-signals of equal length;

[0037] In this embodiment of the invention, the sub-length can be set to 1024.

[0038] Step 3: Encode the sub-signals into a two-dimensional graph domain using a refined recursive graph;

[0039] The parameter settings for the refined recursion graph are shown in Table 1. Then, the calculation process of the refined recursion graph is as follows ( Figure 2 Obtain the corresponding two-dimensional domain. Figure 3 ).

[0040] Table 1. Parameter settings for the refined recursion graph

[0041]

[0042] like Figure 2 As shown, step 3 can be achieved in the following way:

[0043] Assuming a vibration signal corresponds to a time-domain sub-signal set x = {x(i), i = 1, 2, ..., N}, given a time embedding dimension d and a time delay τ, phase space reconstruction of each sub-signal yields the corresponding embedded sample sequence X in the phase space, as shown in the formula:

[0044]

[0045] The similarity distance between all paired embedded sample sequences in the phase space is calculated using Euclidean distance, and the formula is as follows:

[0046]

[0047] Calculate the trend information of the phase space vector:

[0048]

[0049]

[0050] Calculate the refined recursive matrix based on similarity distance and trend information:

[0051] RD i,j =S i,j ×||X i -X j ||

[0052] Encode the refined recursive matrix into a two-dimensional graph domain:

[0053] F = imshow(RD)

[0054] Where x(i) represents the i-th value in a time series x, N represents the length of the time series, L represents the length after phase space reconstruction, τ represents the time delay, and X i (m) represents the m-th element in the i-th phase space vector.

[0055] Step 4: Divide the two-dimensional graph domain into training and validation sets proportionally; divide the dataset into two parts, where 80% is used as the training set to train the deep residual neural network, and 20% is used as the test set to verify the network's classification effect on health status.

[0056] Step 5: Construct an intelligent fault diagnosis model using a ResNet18 network, input the training set for model training; use PyTorch to call the pre-trained ResNet18 network, and change the output category of the classification layer to the actual number of state categories;

[0057] The hyperparameter settings for the ResNet18 neural network are shown in Table 2. The model is trained using this parameter set.

[0058] Table 2 Hyperparameter settings for ResNet18 neural network

[0059]

[0060] For simplicity, the accuracy rate can be calculated using the following formula:

[0061]

[0062] Wherein, TP refers to the number of positive classes predicted as positive, TN refers to the number of positive classes predicted as negative, FP refers to the number of negative classes predicted as positive, and FN refers to the number of negative classes predicted as negative.

[0063] Step 5 can be achieved in the following way:

[0064] 1) Preprocessing: The two-dimensional graph features F are transformed into a size of 224×224 and used as the input to the network;

[0065] 2) A ResNet18 residual neural network is used as the learning network to construct an intelligent fault diagnosis model for wind turbines. The specific implementation process is as follows:

[0066] ResNet18 consists of 17 convolutional layers, 2 pooling layers, and 1 fully connected layer. Except for the first convolutional layer (conv1), the remaining 16 convolutional layers are divided into four residual convolutional blocks: conv2_x, conv3_x, conv4_x, and conv5_x. The computational formula for each convolutional layer is as follows:

[0067]

[0068] Where, x i For the input of the i-th channel, y j Let M be the output of the j-th channel, k be the convolution kernel, b be the bias term, and M be the value of M. j This represents the number of channels.

[0069] Each convolutional block includes two residual units, each consisting of two convolutional layers connected in one step. An identity connection is established between the input and output of the two convolutional layers, specifically:

[0070] x l+1 =x l +F(x l ,k)

[0071] Where, x l+1 and x l Let F represent the output and input of the l-th residual unit, respectively, and F be the residual mapping.

[0072] The ReLU activation function is used to perform a non-linear transformation on each convolutional output y. The specific calculation process is as follows:

[0073] u = max{y, 0}

[0074] Where u is the output obtained by the ReLU function, and y is the output of the convolutional layer.

[0075] Modify the number of outputs of the last fully connected layer to the corresponding number of wind turbine fault types;

[0076] By connecting the above layers in sequence, the structure of the ResNet18 neural network can be represented as: conv1, max pooling layer, 4 sets of residual convolutional blocks, average pooling layer, and fully connected layer.

[0077] 3) Input the training set, select cross-entropy as the loss function, and train the ResNet18 network using gradient descent. The cross-entropy formula is:

[0078]

[0079] Where o is the actual label, y is the predicted label, and Q is the number of fault categories.

[0080] Step 6: Substitute the samples from the test set into the trained ResNet18 to verify their categories.

[0081] The ResNet18 identification results were compared with the actual operation of the wind turbine to calculate the diagnostic accuracy, thus validating the ResNet18 fault identification model. The diagnostic results are shown in Table 3. The method proposed in this invention achieves an accuracy of 90.41%, which is superior to fault diagnosis methods based on 1D-CNN and ResNet18.

[0082] Table 3 Comparison of Diagnostic Results

[0083]

[0084] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0085] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A wind turbine fault diagnosis method integrating graph domain coding and residual neural networks, characterized in that, include: (1) Collect vibration signals of the wind turbine under various conditions. The vibration signals are historical data of different operating conditions recorded by the wind turbine and wind power plant during operation. Set labels according to the condition category. The number of labels is equal to the number of condition categories. Specifically, they include normal state, broken tooth, base crack, tooth damage, gearbox failure and generator failure. (2) Divide each vibration signal into non-overlapping sub-signals of equal length; (3) Reconstruct the phase space of each sub-signal corresponding to each vibration signal, with an embedding dimension d=3 and a time delay τ=1, to obtain the embedded sample sequence in the phase space corresponding to each vibration signal; use Euclidean distance to calculate the similarity distance between all pairs of embedded sample sequences in the phase space; Calculate the trend information of phase space vectors, calculate the fine recursive matrix based on similarity distance and trend information, and encode the fine recursive matrix into a two-dimensional texture map domain; (4) Divide the two-dimensional texture map features into training set and validation set according to the proportion; (5) Transform the size of the two-dimensional texture map domain to 224×224 and use it as the input of the ResNet18 residual neural network. The optimizer is SGD, the initial learning rate is 1e-4, the batch size is 32, and the Maxepoch is 100. Input the training set, select cross-entropy as the loss function and gradient descent to train the ResNet18 network and build an intelligent fault diagnosis model. (6) Input the test set into the trained residual neural network to complete the intelligent fault diagnosis of the wind turbine.

2. The method according to claim 1, characterized in that, Step (4) includes: The two-dimensional texture domain is divided into two parts, with 80% of the data randomly selected as the training set and the remaining 20% ​​as the test set.

3. The method according to claim 2, characterized in that, ResNet18 consists of 17 convolutional layers, 2 pooling layers, and 1 fully connected layer. Except for the first convolutional layer, the remaining 16 convolutional layers are divided into 4 groups of residual convolutional blocks. Each convolutional block includes 2 residual units, each consisting of two convolutional layers connected in one step. An identity connection is established between the input and output of these two convolutional layers, and the ReLU activation function is applied to each convolutional output. y A nonlinear transformation is performed, and the number of outputs of the last fully connected layer is modified to correspond to the number of wind turbine fault types. The above layers are then connected in sequence. The structure of the ResNet18 neural network can be represented as: conv1, max pooling layer, 4 sets of residual convolutional blocks, average pooling layer, and fully connected layer.

4. The method according to claim 3, characterized in that, The test set is input into the trained ResNet18 neural network, and the network output is compared with the actual label category of the wind turbine to obtain the diagnostic accuracy, thus completing the intelligent fault diagnosis of the wind turbine.

Images